Anthropic

Models and offerings #

Who they are

This hub is for engineers who want to call Claude from code, and for anyone comparing hosted providers. Anthropic was founded in 2021 by a group who left OpenAI, is organised as a public benefit corporation, and released the first Claude model in March 2023. Its research output is heavy on safety and interpretability work (the interpretability and safety hubs cover it), and its commercial focus since 2024 has been coding and agentic tool use.

In the field it is the second of the three large closed-model labs by API usage, and the one whose API format diverges most from OpenAI’s. It publishes no parameter counts for any Claude model, and its weights are not available to download anywhere.

Tooling and APIs

pip install anthropic installs the Python SDK, and TypeScript, Java, Go, Ruby, C# and PHP SDKs exist. Keys come from the Console and go in ANTHROPIC_API_KEY. The same models are sold through Amazon Bedrock, Google Cloud (both a Google-managed and an Anthropic-managed offering) and Microsoft Foundry, with slightly different model id strings on each, and the SDK has a client class per platform.

The current lineup at the time of writing, per the vendor’s model page: Claude Fable 5.2 (8 USD per million input tokens, 40 USD output), Claude Opus 5.5 (4 and 20), Claude Sonnet 5.5 (2 and 10) and Claude Haiku 5.5 (0.10 and 0.50 for prompts up to 100 k tokens, five times that above). All four have a 1 M token context window, 128 k output and a stated knowledge cutoff of June 2026. Release dates are not listed on the overview page, which instead commits to retirement dates no sooner than late 2027. Prompt cache reads cost 5 to 10 percent of the input price and the Batch API halves both directions. Parameter counts are undisclosed.

The Messages call is the whole API surface for text:

# needs-api
import os
import anthropic

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])
message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    system="Answer in one sentence.",
    messages=[{"role": "user", "content": "What does a learning rate control?"}],
)
print(message.content[0].text)

Thinking is adaptive on the current models and always on for Opus 5.5 and Fable 5.2, so the older budget_tokens setting is rejected. The effort parameter (low to max) is the control, and Opus 5.5 defaults to medium. Pain points: the tokeniser introduced with the 4.7 generation produces about 30 percent more tokens for the same text than earlier models, which raises bills on migration. A stop_reason of refusal can arrive on a 200 response and must be handled. Prefilling the assistant turn, a common trick on older models, now returns an error.

Migration notes

From OpenAI: there is no base_url trick, the SDK and request shape are different. The system prompt is a top-level system field rather than a message, content is a list of typed blocks (text, image, document, tool_use, tool_result), tool results go back inside a user message, and parallel tool calls must all be answered in the same message. Structured output is output_config.format with a JSON schema. Our prompt patterns reference shows the two formats side by side, and LiteLLM or the OpenAI-compatible endpoints on Bedrock and Foundry are options if you refuse to maintain two code paths.

Prompt caching is explicit: you mark breakpoints with cache_control and keep stable content first. Code ported from a provider with automatic caching often gets a zero cache hit rate until the prompt is reordered.

Between Claude generations: Anthropic publishes a migration guide per model, and the changes are not cosmetic. Moving from the 4.x to the 5.x models removes disabled thinking on some tiers, removes forced tool_choice on others, and ties thinking blocks to the model and conversation that produced them. Re-run your evaluation set before switching a production model id.

Watch for

Lock-in is in the stateful extras rather than the model call: Managed Agents sessions, stored files, Skills and message threads live on Anthropic’s side. Claude Code transcripts are local, which helps.

Rate limits are per organisation and tier and rise with spend, so a fresh key will hit them in a load test. Some models are restricted: the Mythos line is sold only to organisations verified through Anthropic’s programmes, and Fable 5.2 is not available under zero-data-retention terms unless Anthropic agrees. API data is not used for training by default under the commercial terms. Consumer claude.ai conversations have separate settings.

Model retirement is announced with a stated minimum date, and dateless ids on the current generation are pinned snapshots, so claude-opus-5-5 will not silently change under you. Vendor benchmark claims on launch pages are the vendor’s. Confident wrong output (hallucination) remains possible, and the safety classifiers that produce refusals are a product decision you cannot switch off. Regional endpoints on partner clouds carry a 10 percent premium.